A data-driven based fault detection system for grid-connected doubly-fed wind turbines

Abstract

The application discloses a kind of data-driven doubly-fed fan grid-connected fault detection systems, belong to power grid system fault detection field, including sensor detection module, edge computing module, data processing module, central integrated module, fault detection module, and decision module;The application introduces the architecture of distributed data processing, solves the real-time response and high-dimensional data processing problem caused by the complexity of doubly-fed fan system, can significantly reduce the influence of power grid frequency fluctuation on system, reduce power grid pressure and load, improve the response speed and flexibility of system.By multistage processing and edge computing technology, the fault detection function is realized, can detect fault in time effectively, reduce the possibility of fault occurrence, reduce the maintenance and operation cost of system, so as to improve the reliability and economy of overall operation of wind farm.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power grid system fault detection, and in particular to a data-driven doubly-fed wind turbine grid-connected fault detection system. BACKGROUND

[0002] To achieve the "double carbon" goal, China urgently needs to improve the flexibility and stability of wind power grid connection. Doubly-fed wind turbine (DFIG) has become an important choice for wind power development due to its high efficiency and good grid connection characteristics. However, various fault risks faced by DFIG during grid connection, such as voltage fluctuation, frequency anomaly, etc., pose challenges to the safety of its operation and the stability of the power grid. Therefore, it is particularly necessary to improve fault detection and management technology.

[0003] Traditional fault detection methods mainly rely on physical model-based analysis, such as current, voltage feature analysis and control system state feedback, etc. However, these methods may not be able to quickly and accurately detect faults when faced with complex, nonlinear power grid environments, especially for potential hidden faults or complex fault patterns, traditional methods have limitations. SUMMARY

[0004] The purpose of the present application is to solve the defects in the prior art and propose a data-driven doubly-fed wind turbine grid-connected fault detection system.

[0005] In order to achieve the above purpose, the present application adopts the following technical solutions:

[0006] In a first aspect, the present application provides a data-driven doubly-fed wind turbine grid-connected fault detection system, which comprises a sensor detection module, an edge computing module, a data processing module, a central integration module, a fault detection module, and a decision module.

[0007] The sensor detection module collects relevant electrical and mechanical parameters of the wind turbine by sensors distributed at different parts of the system.

[0008] The edge computing module is placed near the sensor location for local rapid preliminary processing and fault diagnosis.

[0009] The specific processing method of the edge computing module is as follows:

[0010] Step S1, the edge computing module first performs noise reduction and filtering processing on the electrical and mechanical parameters collected by the sensor, and removes noise and high-frequency interference in the data through wavelet transform or Kalman filtering technology.

[0011] Step S2, the processed data is subjected to lightweight processing through feature extraction and dimension reduction algorithm, key feature quantities are extracted from it, while redundant information is reduced, and data transmission and calculation load are reduced.

[0012] Step S3, based on the feature data after noise reduction and lightweight processing, the system uses a lightweight machine learning model to perform fault positioning and type prediction, quickly identifies abnormalities, and outputs preliminary fault information;

[0013] The data processing module processes the feature data transmitted back after processing by the edge computing module, and further analyzes the overall system;

[0014] The central integration module integrates the transmission data into a global view;

[0015] The fault detection module is used to analyze complex faults according to the global view and transmit fault information to the decision module;

[0016] The decision module is used for operation and maintenance personnel to monitor system operation state, configuration parameters, and perform operation and adjustment.

[0017] In one possible implementation, the fault detection module has the following specific steps:

[0018] Step V1, according to the data transmitted to the detection module after processing, first perform preliminary anomaly detection, and use threshold detection algorithm to perform preliminary anomaly detection on the processed data, for example, mark voltage instantaneous drop, current surge, etc. as abnormal phenomena; second, based on the pre-set fault judgment rule, quickly locate the possible fault area, match the abnormal parameters with the corresponding equipment or system parts, and preliminarily determine the fault area;

[0019] Step V2, according to the preliminary positioning result, pass it to the upper level processing unit, compare it with the historical data by the historical data comparison module, further confirm whether it is a real fault, if the fault position is not determined or suspected to be a multi-point fault, recheck the related data, perform secondary screening, and finally confirm the fault position;

[0020] Step V3, after confirming the fault position, more detailed feature extraction is performed on the related data of the fault area, and the extracted feature quantity is matched with the pre-defined fault type feature library using neural network, according to the combination of different features, the possible fault type is judged, for example: current anomaly + voltage anomaly can indicate short circuit fault; vibration + temperature anomaly may indicate bearing wear or mechanical jam;

[0021] Step V4, the fault detection module uses multivariate analysis method to analyze the correlation of multiple abnormal parameters, comprehensively judges the fault type, combines the historical data of the fan operation, the equipment characteristics and the fault mode library, judges the severity and influence range of the fault type;

[0022] Step V5, the fault location, fault type, severity, etc. information is summarized and standardized, converted into a format that the system can recognize, and passed to the central control layer through the system bus, with the priority of the fault type and possible operation suggestions marked.

[0023] In one possible implementation, the fault area positioning algorithm mentioned in step V1 is as follows:

[0024] Step V11, the sensor data is monitored through a threshold detection algorithm, and by presetting the normal range of multiple parameters, it is monitored whether the real-time data exceeds these ranges, and the abnormal phenomenon is preliminarily diagnosed. For grid-connected detection of a doubly-fed wind turbine system, the electrical parameters monitored here are as follows:

[0025] X(t) = {I(t), U(t), P(t)}, t = 0, 1, 2, … T

[0026] In the formula, I(t) is the current of the wind turbine at time t, U(t) is the voltage at that time, and P(t) is the power at that time.

[0027] When the monitored electrical parameters exceed the preset threshold range, the system will mark this moment as abnormal. The threshold monitoring rule is as follows:

[0028]

[0029] Step V12, in order to avoid the fact that in some complex systems, an abnormal parameter may cause multiple regions to perform, Bayesian inference is introduced to calculate the probability of fault occurrence according to historical fault data and current detection data, and the Bayesian formula is used to detect abnormal data to update the fault region probability as follows:

[0030]

[0031] In the formula, P(R i ) is the initial probability of each region occurring a fault according to historical data, where R i represents region i, and P(R i |data) is the probability of observing the current data when the region R i fails.

[0032] In one possible implementation, the fault type analysis algorithm mentioned in step V3 is as follows:

[0033] Step V31, by analyzing the electrical and mechanical signals collected during the operation of the doubly-fed wind turbine, feature variables that can represent the system state and potential faults are extracted; electrical signals include current and voltage, etc., which reflect the electrical operation state of the wind turbine. By extracting the time domain and frequency domain features of these signals, electrical faults can be identified.

[0034] In the case of short circuit or overload, the instantaneous current fluctuates abnormally, the maximum instantaneous value of the current signal is extracted, and the formula is as follows:

[0035]

[0036] In the formula, I peak represents the maximum value of the current extracted in the entire time sequence;

[0037] The mean value can reflect the overall current level of the system, and is used to capture the current change trend under long-time operation, and the formula is as follows:

[0038]

[0039] In the formula, T is the sampling period, and I(t) is the current signal;

[0040] In order to extract the frequency components in the current signal, especially the high frequency components, the electrical harmonic interference or instantaneous fault in the system can be identified, and the Fourier Transform is used to convert the time domain signal to the frequency domain, as follows:

[0041]

[0042] In the formula, I(t) represents the intensity of the current signal at different frequencies f;

[0043] The mechanical signal includes vibration and temperature data, which can reflect the mechanical state of the fan;

[0044] Step V32, after feature extraction, the feature quantity is matched with the predetermined fault type feature library, and in the case of complex and nonlinear problems, the extracted features are matched with the predefined fault types by using a neural network;

[0045] First, the extracted multi-dimensional features are input into the neural network, each feature quantity is taken as an input node of the network, and the activation value of the first layer of neurons is generated by combining the input features through the weight matrix W and the bias term b, and the formula is as follows:

[0046] z1=W1·F+b1

[0047] In the formula, F=[I peak ,U drop ,V amp ,T rate ] is the input feature quantity, W1 is the weight matrix, and b1 is the bias quantity;

[0048] Secondly, the input data is nonlinearly mapped by using the activation function ReLU, the positive value is retained, the negative value is filtered, and the network helps to extract the complex features related to the fault type layer by layer, and the expression is:

[0049] ReLU(z) = max(0, z)

[0050] Combining them together forms a multi-layer perceptron that can extract features, the specific process is represented as:

[0051] z = W - ReLU(z) + b

[0052] Step V33, after the activation and weight update of each layer, the network gradually extracts patterns related to fault types, each layer of neurons generates the input of the next layer according to the output of the previous layer, until the network reaches the output node of the last layer, and the Softmax classifier is used in the last layer to convert the output layer result into the probability of different fault types, the expression is:

[0053]

[0054] In the formula, P(y = i | z) represents the probability of fault type i, and zi is the output of the last layer of the neural network; according to the output, the fault type with the highest probability is selected as the final output.

[0055] In one possible implementation, the multivariate analysis algorithm proposed in step V4 is as follows:

[0056] Step V41, in this stage, we have obtained the multi-dimensional input data of feature extraction, and further analyzed the relationship between these features through the multi-dimensional classification method of SVM to determine the fault type; SVM constructs an optimal hyperplane to distinguish different fault categories, which can map different feature combinations to different fault types, and the mathematical description is as follows:

[0057] f(X) = sign(wTX + b)

[0058] In the formula, w is the model weight vector, b is the bias value, and X = [I peak ,U drop ,V amp ,T rate ];

[0059] Step V42, the system adjusts w and b according to the training data to ensure that the input feature data X is correctly classified as a certain fault type, for example: peak and U drop abnormal at the same time, the system will determine that this case is most likely a short circuit fault through SVM; V amp and T rate abnormal at the same time, the system may classify it as a mechanical fault, such as bearing wear.

[0060] Step V43, by comparing the abnormality degree between different features, the SVM can distinguish the severity of the fault, for a given data point xi, the distance from the hyperplane is calculated as follows:

[0061]

[0062] where w T x i +b represents the projection result of the feature point to the hyperplane, ||w|| is the norm of the weight vector, used for normalization; the greater the distance d(x i ), the farther the data point is from the classification boundary, indicating that the severity of the fault is higher, after the final fault determination by the SVM, the system generates a detailed fault report. The report includes: fault type, fault severity, potential impact range. BRIEF DESCRIPTION OF DRAWINGS

[0063] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, for explaining the application, and do not constitute a limitation on the application.

[0064] Figure 1 A block diagram of a data-driven DFIG grid-connected fault detection system is proposed for the application;

[0065] Figure 2 A flowchart of a data-driven DFIG grid-connected fault detection module is proposed for the application. DETAILED DESCRIPTION

[0066] In order to better understand the technical scheme of the present application, the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0067] An embodiment of a data-driven DFIG grid-connected fault detection system applied to electrical fault detection, this embodiment shows the fault detection process of the system when the fan electrical operation is abnormal; first, the sensor detection module collects current, voltage, power and other key electrical parameters, the edge computing module denoises and filters the collected data, extracts features and performs lightweight processing; then, the system uses a lightweight machine learning model to quickly identify current transient abnormal fluctuations and voltage drop phenomena, and locates them as potential short circuit faults; the fault detection module further analyzes the feature data, combines historical data and feature library to confirm the possibility of short circuit fault, and quantifies the severity of the fault through SVM multivariate analysis; finally, the decision module transmits the detected short circuit fault position, severity and other information to the operation and maintenance personnel, and provides priority and emergency treatment suggestions; it should be clear that the embodiments are only part of the embodiments of the present application, and other unlisted embodiments are also within the scope of the present application.

[0068] Figure 1 A block diagram of a data-driven DFIG wind turbine grid-connected fault detection system is provided in the present application, as shown in the figure. Figure 1 The system comprises a sensor detection module, an edge computing module, a data processing module, a central integration module, a fault detection module, and a decision module.

[0069] The sensor detection module collects relevant electrical and mechanical parameters of the wind turbine by sensors distributed in different parts of the system.

[0070] The edge computing module is placed near the sensor location for local rapid preliminary processing and fault diagnosis.

[0071] The specific processing method of the edge computing module is as follows:

[0072] Step S1, the edge computing module first performs noise reduction and filtering processing on the electrical and mechanical parameters collected by the sensor, and removes noise and high-frequency interference in the data through wavelet transform or Kalman filtering technology.

[0073] Step S2, the processed data is processed by feature extraction and dimension reduction algorithm, and the key feature quantity is extracted, while the redundant information is reduced, and the data transmission and calculation load are reduced.

[0074] Step S3, based on the feature data after noise reduction and lightweight processing, the system uses a lightweight machine learning model to perform fault positioning and type prediction, quickly identifies abnormalities, and outputs preliminary fault information.

[0075] The data processing module processes the feature data transmitted back by the edge computing module, and further analyzes the overall system.

[0076] The central integration module integrates the transmission data into a global view.

[0077] The fault detection module is used to analyze complex faults according to the global view and transmit fault information to the decision module.

[0078] The decision module is used for operation and maintenance personnel to monitor the system running state, configure parameters, and perform operation and adjustment.

[0079] Figure 2 A flow chart of a data-driven DFIG wind turbine grid-connected fault detection module is provided in the present application, as shown in the figure. Figure 2As shown, the module transmits the data processed by the integrated module to the fault detection module by collecting the data, first identifies the abnormal area by using the threshold detection algorithm, secondly updates the probability according to the Bayesian inference based on the historical data provided by the system, and finally identifies and confirms the fault area; after identifying the fault area, the electrical and mechanical signals of the fault area are extracted, matched with the fault type, and combined with the electrical and mechanical multi-dimensional features in the grid-connected operation state for multi-variable in-depth analysis to determine the fault type, position and depth, and generate a fault report.

[0080] In the embodiment of the application, the specific steps of fault detection of the fault detection module are as follows:

[0081] Step V1, according to the data transmitted to the detection module after processing, first, the processed data is subjected to preliminary abnormality detection by a threshold detection algorithm, for example, transient voltage drop, sudden current increase, etc. are marked as abnormal phenomena; secondly, based on the pre-set fault judgment rule, the possible fault area is quickly located, the abnormal parameters are matched with the corresponding equipment or system parts, and the area where the fault occurs is preliminarily determined;

[0082] Step V2, according to the result of preliminary positioning, the result is transmitted to the upper level processing unit, and the historical data are compared with the historical data by the historical data comparison module to further confirm whether the fault is real or not, if the fault position is uncertain or suspected to be a multi-point fault, the related data is re-verified for secondary screening, and finally the fault position is confirmed;

[0083] Step V3, after confirming the fault position, more detailed feature extraction is performed on the related data of the fault area, the extracted feature quantity is matched with the pre-defined fault type feature library using a neural network, and the possible fault type is judged according to the combination of different features, for example: current abnormality + voltage abnormality can indicate short circuit fault; vibration + temperature abnormality can indicate bearing wear or mechanical jamming;

[0084] Step V4, the fault detection module uses a multi-variable analysis method to analyze the correlation of multiple abnormal parameters, comprehensively judges the fault type, and judges the severity and influence range of the fault type in combination with the historical data of the fan operation, the equipment characteristics and the fault mode library;

[0085] Step V5, the information of fault position, fault type, severity, etc. is summarized and standardized, converted into a format that can be recognized by the system, transmitted to the central control layer through the system bus, and the priority of the fault type and the possible operation suggestion are marked.

[0086] In the embodiment of the application, the fault area positioning algorithm in step V1 is as follows:

[0087] Step V11, monitor the sensor data by threshold detection algorithm, monitor whether the real-time data is out of the normal range of the preset multiple parameters, preliminarily diagnose the abnormal phenomenon, and monitor the electrical parameters as follows for the grid-connected detection of the doubly-fed wind turbine system:

[0088] X(t)={I(t),U(t),P(t)},t=0,1,2,….T

[0089] In the formula, I(t) is the current of the wind turbine at time t, U(t) is the voltage at the moment, and P(t) is the power at the moment.

[0090] When the monitored electrical parameters are out of the preset threshold range, the system will mark this moment as abnormal. The threshold monitoring rule is as follows:

[0091]

[0092] Step V12, in order to avoid that in some complex systems, one parameter abnormality may cause multiple regions to perform parameters, Bayes inference is introduced to calculate the fault occurrence probability according to historical fault data and current detection data, and Bayes formula is used to detect abnormal data to update the fault region probability as follows:

[0093]

[0094] In the formula, P(R i ) is the initial probability of each region occurring fault according to historical data, wherein R i represents region i, and P(R i |data) is the probability of observing current data when region R i occurs fault.

[0095] In the embodiment of the application, the fault type analysis algorithm proposed in step V3 is as follows:

[0096] Step V31, by analyzing the electrical and mechanical signals collected during the operation of the doubly-fed wind turbine, feature variables capable of representing the system state and potential faults are extracted; the electrical signals include current and voltage and the like, which reflect the electrical operation state of the wind turbine, and by extracting the time domain and frequency domain features of these signals, the electrical faults can be identified;

[0097] In the case of short circuit or overload, the instantaneous current fluctuates abnormally, the maximum instantaneous value of the current signal is extracted, and the formula is as follows:

[0098]

[0099] In the formula, I peak represents the maximum value of the current extracted in the entire time sequence;

[0100] The average value can reflect the overall current level of the system, and is used to capture the current trend over a long period of time, and the formula is as follows:

[0101]

[0102] In the formula, T is the sampling period, and I(t) is the current signal;

[0103] In order to extract the frequency components in the current signal, especially the high frequency components, to identify the electrical harmonic interference or transient fault in the system, the Fourier Transform is used to convert the time domain signal to the frequency domain, as follows:

[0104]

[0105] In the formula, I(t) represents the intensity of the current signal at different frequencies f;

[0106] The mechanical signal includes vibration and temperature data, which can reflect the mechanical state of the fan;

[0107] After feature extraction, the feature quantity is matched with the predetermined fault type feature library. When dealing with complex and nonlinear problems, the extracted features are matched with the predefined fault types using a neural network;

[0108] First, the extracted multi-dimensional features are input into the neural network, and each feature quantity is used as an input node of the network. Through the weight matrix W and the bias term b, the activation value of the first layer of neurons is generated by combining the input features, and the formula is as follows:

[0109] z1=W1·F+b1

[0110] In the formula, F=[I peak ,U drop ,V amp ,T rate ] is the input feature quantity, W1 is the weight matrix, and b1 is the bias quantity;

[0111] Second, the input data is nonlinearly mapped using the activation function ReLU, which retains positive values and filters negative values, helping the network to extract complex features related to fault types layer by layer, and the expression is:

[0112] ReLU(z)=max(0,z)

[0113] Combining them constitutes a multi-layer perceptron that can extract features, and the specific process is represented as:

[0114] z=W·ReLU(z)+b

[0115] Step V33, after the activation and weight update of each layer, the network gradually refines the pattern related to the fault type, each layer of neurons generates the input of the next layer according to the output of the previous layer, until the network reaches the output node of the last layer, and the Softmax classifier is used in the last layer to convert the output layer result into the probability of different fault types, the expression is:

[0116]

[0117] In the formula, P(y=i|z) represents the probability of fault type i, and zi is the output of the last layer of the neural network; according to the output, the fault type with the highest probability is selected as the final output.

[0118] In the embodiment of the application, the multivariate analysis algorithm proposed in step V4 is as follows:

[0119] Step V41, in this stage, the multi-dimensional input data of feature extraction has been obtained, and the relationship between the features is further analyzed by the multi-dimensional classification method of SVM to determine the fault type; the SVM distinguishes different fault categories by constructing an optimal hyperplane, which can map different feature combinations into different fault types, and the mathematical description is as follows:

[0120] f(X)=sign(wTX+b)

[0121] In the formula, w is the model weight vector, b is the bias value, X=[I peak ,U drop ,V amp ,T rate ];

[0122] Step V42, the system adjusts w and b according to the training data to ensure that the input feature data X is correctly classified into a certain specific fault type, for example: I peak and U drop abnormal at the same time, the system will determine that this situation is most likely a short circuit fault through SVM; V amp and T rate abnormal at the same time, the system may classify it as a mechanical fault, such as bearing wear.

[0123] Step V43, by comparing the abnormality degree between different features, the SVM can distinguish the urgency of the fault, and for a given data point xi, the distance from the hyperplane is calculated as follows:

[0124]

[0125] In the formula, w T x i+b represents the projection result of feature points to the hyperplane, ||w|| is the norm of the weight vector, which is used for normalization; the greater the distance d(x i ) is, the farther the data point is from the classification boundary, indicating that the severity of the fault is higher. After the final fault determination by the SVM, the system generates a detailed fault report. The report includes: fault type, fault severity, potential impact range.

Claims

1. A data-driven grid-connected fault detection system for doubly-fed wind turbines, characterized in that, It includes a sensor detection module, an edge computing module, a data processing module, a central integration module, a fault detection module, and a decision-making module; The sensor detection module consists of sensors distributed in different parts of the system, used to collect relevant electrical and mechanical parameters of the wind turbine; The edge computing module is placed close to the sensor for rapid local preliminary processing and fault diagnosis; The specific processing method of the edge computing module is as follows: Step S1: The edge computing module first performs noise reduction and filtering on the electrical and mechanical parameters collected by the sensors. In response to the common harmonic interference and high-frequency oscillation problems in grid-connected systems, and combined with the power electronic characteristics of the doubly fed wind turbine, an improved Kalman filter algorithm and multi-scale wavelet transform are used to remove noise. Step S2: The processed data is lightweighted using feature extraction and dimensionality reduction algorithms to extract key features. Step S3: Based on the feature data after noise reduction and lightweight processing, the system adopts a lightweight machine learning model and combines it with the dynamic change trend of grid connection parameters to achieve preliminary fault location and type prediction. The data processing module processes the feature data returned by the edge computing module and further analyzes the system as a whole. The central integration module integrates the transmitted data into a global view; The fault detection module is used to analyze complex faults based on a global view and transmit the fault information to the decision-making module. The specific processing steps of the fault detection module are as follows: Step V1: Based on the processed data transmitted to the detection module, use the dynamic threshold detection algorithm to perform preliminary anomaly detection, especially for fluctuations in grid-connected voltage and current. Combined with preset fault judgment rules, quickly locate the area where the fault occurred and associate it with the corresponding system components. Step V2: After initial location, further confirm whether the fault actually occurred by comparing with historical data; If the fault location is unclear or multiple faults are suspected, the data should be cross-validated and filtered again to finally confirm the fault location and reduce the chance of misjudgment. Step V3: After confirming the fault location, the system performs more detailed feature extraction. Combining the power electronic and mechanical characteristics of the doubly fed wind turbine, the system extracts feature quantities and performs classification and matching through a neural network intelligent algorithm, matching the fault type with the preset feature library. Step V4: The fault detection module uses support vector machine to perform multivariate analysis, combining the electrical and mechanical multidimensional characteristics under grid-connected operation to determine the fault type and its severity, and generate a comprehensive report to provide accurate fault nature and potential impact range. Step V5, fault type, location, and severity information are transmitted to the central control layer after standardized processing, with the priority of the fault type and operation suggestions marked to help maintenance personnel make decisions; The decision-making module is used by operations and maintenance personnel to monitor the system's operating status and configuration parameters, and to perform operations and adjustments.

2. The data-driven doubly-fed wind turbine grid connection fault detection system according to claim 1, characterized in that, The fault location algorithm proposed in step V1 is as follows: Step V11: Real-time monitoring of grid-connected electrical parameters is performed using a threshold detection algorithm. When the monitored electrical parameters exceed the preset threshold range, the system will mark the current time as abnormal. Step V12: To avoid misdiagnosis caused by multiple parameter anomalies, Bayesian inference combined with historical data is used to calculate the probability of fault occurrence and locate the fault area, thereby improving detection accuracy.

3. The data-driven doubly-fed wind turbine grid connection fault detection system according to claim 1, characterized in that, The fault type analysis algorithm proposed in step V3 is as follows: Step V31: Extract time-domain and frequency-domain features from the electrical and mechanical signals during the grid connection process of the doubly fed wind turbine, and combine Fourier transform to analyze the harmonic components in the electrical signals to reflect the instantaneous fault characteristics of the system. Step V32: After feature extraction, the extracted multidimensional features are classified by a neural network. The extracted multidimensional features are input into the neural network, with each feature serving as an input node. The activation value of the first layer of neurons is generated by combining the input features with the weight matrix W and the bias term b. Secondly, the ReLU activation function is used to perform non-linear mapping on the input data, retaining positive values ​​and filtering negative values, which helps the network to extract complex features related to the fault type layer by layer. Combining these features constitutes a multilayer perceptron that can extract features. Step V33: After activation and weight update of each layer, the network gradually extracts patterns related to fault types. Each neuron generates the input of the next layer based on the output of the previous layer until the network reaches the output node of the last layer. In the last layer, the Softmax classifier is used to convert the output layer results into probabilities of different fault types. The fault type with the highest probability is selected as the final output based on the output.

4. The data-driven doubly-fed wind turbine grid connection fault detection system according to claim 1, characterized in that, The multivariate analysis algorithm proposed in step V4 is as follows: Step V41: Build a model using SVM and generate the optimal hyperplane through training data. This hyperplane is used to distinguish different types of faults in the feature space, ensuring that different fault categories can be accurately separated in high-dimensional space. Step V42: SVM maps the input data points into the feature space according to different combinations of features, and classifies these data points according to the optimal hyperplane; Step V43: SVM not only classifies fault types, but also assesses the severity of faults by calculating the distance of feature data relative to the hyperplane. Based on the SVM classification and assessment results, a detailed fault report is generated.

Images